Align organization card and launch styled IRIS overview

README.md

@@ -1,75 +1,53 @@
 ---
-title: Saccadic | IRIS Training Diagnostics
-emoji: 🌸
+title: Saccadic | IRIS Training Observations
+emoji: 👁
 colorFrom: purple
 colorTo: blue
 sdk: static
+app_file: index.html
 pinned: false
 short_description: A closer look inside neural network training with IRIS
 ---
 
-# SACCADIC
-## IRIS™ · A closer look at AI training
+<img src="https://cdn-avatars.huggingface.co/v1/production/uploads/69d6d16790de87a3b534bf99/SkapdHGpRMem-dCvIvgMx.png" alt="Saccadic purple lotus with a blue eye" width="110">
 
-**A better view while the run is still underway.**
+# IRIS™ · A closer look at AI training
 
-Saccadic LLC develops instruments for understanding neural network training. Our **Intrinsic Rank Inspection System (IRIS)** adds a measured view of internal structure, specialization, and coordination alongside loss and validation curves.
+*A better view while the run is still underway.*
 
-[Explore IRIS](https://www.saccadic.ai/iris) · [Visit Saccadic](https://www.saccadic.ai/) · [Discuss an evaluation](mailto:info@saccadic.ai)
+Saccadic LLC develops instruments for observing neural network training. IRIS, the Intrinsic Rank Inspection System, adds a measured view of internal structure, specialization, and coordination alongside loss and validation curves.
 
----
-
-### The scoreboard tells you the outcome. IRIS helps you inspect the play.
-
-A loss curve can show a change without locating it inside the model. IRIS helps researchers and engineers inspect where representations concentrate, which subjects reach each expert, how domains share or conflict, and how the model's organization changes over time.
-
-The goal is a more focused investigation: connect an observation to a testable question, compare it with task results, and design the next experiment.
-
-### A training movie, built from measurements
-
-| Step | What happens |
-| --- | --- |
-| **Capture** | The tracker takes measurements at fixed training intervals or adjusts spacing to organizational activity, using consistent probe inputs. |
-| **Record** | Measurements accumulate in a self-contained **.iris** file. Training checkpoints keep their usual role. |
-| **Inspect** | The locally running **Bench** reads the growing record during training and opens it for review afterward, without loading the model. |
+[Explore the styled IRIS Space](https://huggingface.co/spaces/SaccadicAI/README) · [IRIS product details](https://www.saccadic.ai/iris) · [Discuss an evaluation](mailto:info@saccadic.ai)
 
-### Look inside the run
+## Look inside the run
 
 - **Flow Map:** inspect measured connections between layers and experts for different input domains.
-- **Representation structure:** track independent directions carrying 95% of activation variation and how dominant directions move.
+- **Representation structure:** observe how activation variation concentrates and how dominant directions move.
 - **Expert routing:** compare which subjects reach which experts, separately for each layer.
-- **Domain interaction:** inspect sharing and training-gradient agreement across subjects.
-- **Training kinetics:** compare coherent weight motion with motion that cancels.
+- **Domain interaction:** inspect sharing and training-gradient agreement alongside loss and task results.
 
-The October 2026 product reference includes **103 built-in diagnostic views across eight scientific tabs**: Domains, Layers, Heads, Experts, Facets, Whole Model, Flow Map, and Concept Space. These are counted views, not independent scalar measures; coverage varies by model, layer, head, and expert.
+## Capture. Record. Inspect.
 
-Each graph has an Explain guide. Teams can report their own labeled scalar signals, derive formulas from captured channels, and export custom charts as PNG images, Excel workbooks, or JSON data.
+The tracker captures measurements using consistent probe inputs. A self-contained **.iris** file records the run. The local **Bench** reads that growing record during training and opens it afterward without loading the model.
 
----
-
-### From a measured run: five operations, different learning times
-
-![Training and held-out accuracy for addition, subtraction, multiplication, division, and exponentiation modulo 113](https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/LsCSdafjN_zw5hphoWKHa.png)
+Each graph has an Explain guide. Teams can define labeled scalar signals, derive formulas from captured channels, and export custom charts.
 
-In this small mixture-of-experts transformer run, multiplication and division reached 90% held-out accuracy before exponentiation, subtraction, and addition. Overall accuracy eventually flattened, while internal measurements continued to change.
+## A measured example
 
-[Read the experiment: Accuracy plateaued. The model kept changing.](https://huggingface.co/blog/Saccadic-AI/accuracy-plateaued-the-model-kept-changing)
+![Training and held-out accuracy by mathematical operation](https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/LsCSdafjN_zw5hphoWKHa.png)
 
-### Fits the customer's training environment
+*Figure 4.4. Training and held-out accuracy for the mathematical operations in this measured run. [Read the experiment and its conditions.](https://huggingface.co/blog/Saccadic-AI/accuracy-plateaued-the-model-kept-changing)*
 
-IRIS fits existing **PyTorch training loops**, with support for GPT / LLaMA / Mistral, BERT-family encoders, and supported mixture-of-experts models. User-defined architectures use a measurement adapter to declare capture points. Capture and inspection can run on **CPU, Apple MPS, or NVIDIA CUDA**.
+## Work in your own environment
 
-Model weights, activations, and probe text are not sent to Saccadic. Licensing and metering contact Saccadic; optional configured relays send selected scalar measurements to experiment trackers including **Weights & Biases, TensorBoard, MLflow, and Aim**.
+IRIS fits existing PyTorch training loops, with capture and inspection on CPU, Apple MPS, or NVIDIA CUDA. Coverage depends on model layout; shared-weight configurations are currently unsupported.
 
-Coverage depends on model layout. Shared-weight configurations are currently unsupported.
+Model weights, activations, and probe text are not sent to Saccadic. Licensing and metering contact Saccadic. Optional configured relays send selected scalar measurements to existing experiment trackers.
 
-### Start with one model, one run, one investigation
+## One model. One run. One concrete investigation.
 
-We welcome conversations with research groups and teams training or fine-tuning proprietary models. A joint evaluation can test whether the readings focus a real investigation and measure overhead in the customer's workload. Academic researchers can request credit grants; requests are reviewed individually.
+We welcome research and joint evaluations. Academic researchers can request credit grants; requests are reviewed individually. See the public product site for availability and licensing terms.
 
-**Contact:** [info@saccadic.ai](mailto:info@saccadic.ai)  
-**Product:** [www.saccadic.ai/iris](https://www.saccadic.ai/iris)
-
----
+[info@saccadic.ai](mailto:info@saccadic.ai) · [www.saccadic.ai/iris](https://www.saccadic.ai/iris)
 
-*Based on the Saccadic IRIS product presentation dated 24 October 2026. IRIS is a measurement instrument; readings should be interpreted alongside gradients, loss, validation, and task results.*
+*IRIS pre-alpha. Product overview based on the presentation dated 24 October 2026. IRIS records observations; interpret readings alongside gradients, loss, validation, and task results.*